The existing HDR image display systems are basically constructed with a combina-tion of two layers. Layer that provides high resolucombina-tion and low luminance placed in front of the layer that provides low resolution and high luminances. Using these layers one after the other provides both high luminance levels and contrast ratio. Contrast ratio is measured as multiplication of two systems. Back-mounted layer can give the desired luminance levels and front-mounted layer can give local contrasts. [29, 30].
Figure 3.4: LED positions for SIM2 display
SIM2 monitors are produced with that principle. In the perceptual experiments SIM2 HDR47 monitor is used. In background, there are 2202 LEDs are placed diagonally.
Each of these LEDs are adjusted independently. Figure 3.4 illustrates the LED po-sitions on the SIM2 display. These LEDs provide high luminance levels. In the foreground there is an LCD layer with 1920x1080 full HD resolution.
SIM2 HDR47 monitor has two modes to display HDR images. One of these modes is the automatic rendering that display contains. The other mode is DVI+ mode, which
takes LED and LCD values as inputs. In order to display the most accurate image with SIM2 HDR47 monitor, first the proposed algorithm in [31] is used.
Figure 3.5: PSF for LEDs
Each LED in SIM2 HDR47 monitor has its own Point Spread Function (PSF) as shown in Figure 3.5. Each LED’s radiance is a cumulative summation with its neigh-bour LEDs’ radiance values. Algorithm described in [31] presents a solution for calculating the cumulative radiance of LEDs to give the desired luminance level and for finding the resulting LCD values. The main stages of this algorithm is as follows;
• Preprocessing
• Finding target backlight
• Iterative scaling
• Finding LCD values
As presented in Figure 3.6, this algorithm [31] gives an halo effect around the local noise patterns. The halo effect is a form of cognitive bias which causes one part to make the whole seem more attractive or desirable. For the perceptual experiments, making the noise more attractive is an unwanted situation. Since the question to the subject during the perceptual tests is whether a noise is noticed or not, there shouldn’t be any eye catcher in the image.
Figure 3.6: Halo effect around the local noise pattern formed with frequency (3,3), DC luminance level 1000 cd/m2, and amplitude 150.
For experimental purposes background luminance should be same in the whole im-age. For that purpose, the proposed algorithm in [31] is revised. Target backlight shouldn’t be calculated but set to desired value. In order to find the desired backlight the following steps are done.
• While keeping the LCD values at 255, LED values are changed from 1 to 255.
• Luminance values of SIM2 HDR47 monitor are saved for each LED value.
• Search table is constructed with the obtained values.
Figure 3.7 illustrates the measured luminance values as a function of the LED values.
LCD values are kept at 255 during the measurement. It is observed that the graph shows a linear increase until the LED values are approximately 100. It then proceeds in a linear manner. The reason for this graph is that led energies reach the maximum of the monitor after a point.
The step for finding LCD values is performed as defined in algorithm [31]. Halo effects are eliminated with the proposed approach. Sample images with the proposed algorithm are provided in Figure 3.8.
Figure 3.7: LED values vs. Luminance graph
Figure 3.8: Local noise pattern without halo effect formed with frequency (3,3), DC luminance level 1000 cd/m2, and amplitude level 150.
CHAPTER 4
PROPOSED METHODOLOGY FOR PSYCHO-VISUAL EXPERIMENTS
Figure 4.1: The overall experimental procedure
QUEST method, which is introduced by Watson and Pelli [32], a procedure that es-timates the ideal threshold in a given number of sequential trials. QUEST method is generally used at psycho-visual experiments where subject is tested. According to the answers of the subject, probability density function for the estimated quantiza-tion level is recalculated [33]. In each trial of the experiment, the subject is asked a question. With each iteration, more information is obtained from the subject and thus more accurate estimations can be done. The experimental methodology, which is proposed by this thesis, also uses QUEST method.
QUEST method is provided in MATLAB environment by Psychtoolbox. However, in order to use this method, a prior threshold estimate and a standard deviation, which is assigned to that estimate, are required. With the purpose of estimating a threshold and a standard deviation, another experimental set up is implemented. In that experi-mental setup, geometric search is used. Figure 4.1 illustrates the whole experiexperi-mental setup.
The images for the DCT quantization noise, as described in Chapter 3 (Construc-tion And Rendering Of DCT Patterns), are used in the experiments. For four
differ-ent luminance levels (100, 500, 1000, 1500 cd/m2) for each of the 30 frequencies ((0,0), (0,1), (0,2), (0,3), (0,5), (0,7), (1,0), (1,1), (1,2), (1,3), (1,5), (1,7), (2,0), (2,1), (2,2), (3,0), (3,1), (3,2), (3,3), (3,5), (3,7), (5,0), (5,1), (5,2), (5,3), (5,5), (5,7), (7,0), (7,1), (7,2), (7,3), (7,5), (7,7)) the images for DCT quantization noise are formed.
More specifically, the images are generated with an amplitude level from 1 to 70 for luminance level 100cd/m2. Accordingly, the images are generated up to an ampli-tude level 380 for 500cd/m2 , whereas up to 500 for 1000cd/m2, and up to 700 for 1500cd/m2.
4.1 Geometric Search
Figure 4.2: Psycho-visual experimental procedure with the geometric search
The aim of the experimental procedure with the geometric search is to determine an initial guess and a standard deviation for that guess for each DCT frequency pattern, which would be given as inputs to the QUEST procedure. Two alternative forced choice procedure [34] is used in this experiment. Two consecutive images are dis-played, where one contains the test stimulus and the other one contains gray level base image. These images are displayed in a random order. Displaying two images and receiving the answer forms one trial. The amplitude of the DCT pattern to be displayed after each iteration is calculated by geometric search.
This experiment is completed for luminance levels 100,500,1000 and 1500 cd/m2 and for each DCT frequency level, with three subjects.
The main stages of the geometric search are described as follows:
• Experiment starts with the highest amplitude of the given DCT pattern.
• Test stimulus which contains the noise and the base image which only contains gray level image are displayed consecutively in a random order for 1 second.
Between two images there is a 0.25 second duration time.
• Subject is asked if any difference is observed between two images.
• Subject presses ’y’ from the keyboard if any difference is observed. Otherwise subject presses ’n’.
• Next trial does not begin until a response is given.
• If the answer is yes then quantization value is decreased, if the answer is no then quantization value is increased.
• The amount of change in the quantization value is calculated as the half of the difference between subject’s latest two trial’s quantization values.
• When this difference is lower than a certain threshold, experiment is terminated.
The final Q value is considered as that subject’s just noticeable quantization level for that experiment set.
Results of two subjects are used to calculate initial quantization value and standard deviation which is provided as inputs for Quest experiment sets. Initial guess is calcu-lated as average of two subject’s just noticeable quantization level for the given DCT pattern set. Standard deviation is calculated with respect to subjects’s just noticeable quantization levels and their average.